An opionated set of convensions I stick to in my Bevy projects.
Up to date as of Bevy 0.11.
All entities must be spawned with a Name and cleanup component at the front of the bundle.
Names assist with debugging. Cleanup components indicate to which state the entity belongs, and to remove it upon exit of that state.
By always having these two "meta" components at the front it makes it easy to spot entities where they are missing.
commands
.spawn((
Name::new("Player"),
cleanup::CleanupInGamePlayingExit,
...
))You can read more about the cleanup pattern I'm using in the bevy cheatbook.
All systems added to the Update schedule should be bound by run conditions on State and SystemSet.
Run states enables easy enabling/disbling of groups of behaviour & reduces systems running when they don't need to.
For example, changing PlayingState to PlayingState::Paused will automatically disable all systems that progress the game and enable systems handle actions related to the pause menu.
System sets force coarse grained ordering leading to predictable behaviour between different parts of the game.
There can be exceptions to this, for example you may have background music or UI animations that should continue in both Playing and Paused.
.add_systems(
Update,
(handle_player_input, camera_view_update)
.chain()
.run_if(in_state(PlayingState::Playing))
.run_if(in_state(GameState::InGame))
.in_set(UpdateSet::Player),
)State transitions should have setup and cleanup specific systems. Their OnEnter and OnExit registration should be co-located.
This means it's easy to see the setup systems and that it has a cleanup system to run.
.add_systems(OnEnter(GameState::MainMenu), main_menu_setup)
.add_systems(OnExit(GameState::MainMenu), cleanup_system::<CleanupMenuClose>)
.add_systems(Update, (...).run_if(in_state(GameState::MainMenu)))Use Events to structure logic and communication between subsystems of your game.
Events allow different parts of your game to opt-in to information they need, and prevents tight coupling.
EventWriters and EventReaders are implemented as a thin layer over a vanilla Vec so they're cheap to use, and subsequent sends will re-use the allocated capacity. Systems that read from the same event can also run in parallel as EventReaders are local to the system.
Here's an example of one way you might structure a projectile hitting an enemy, all the way to audio, visual effects, and achievements being updated. You should evaluate the use of events on a case by case basis, as they're not free, and for simple local operations it can be enough to mutate within the same system.
flowchart TD
Y[Send EnemyHit event]
EnemyHitEvent["EnemyHitEvent
- player
- enemy
- damage"]
ProjectileHitEvent["ProjectileHitEvent
- projectile
- enemy"]
subgraph physics
X[Projectile intersects Enemy] -.-> Y["EventWriter::#60;ProjectileHitEvent#62;::send()"]
end
Y --> ProjectileHitEvent
subgraph gameplay
hit_recv[Handle Projectile Hit]
Z["EventWriter::#60;EnemyHitEvent#62;::send()"]
end
ProjectileHitEvent --> hit_recv
hit_recv --> Z --> EnemyHitEvent
subgraph audio
audio_recv[Play Hit SFX]
end
subgraph vfx
vfx_recv[Play Hit VFX]
end
subgraph achievement
achievement_recv[Damage Total]
end
EnemyHitEvent --> audio_recv & vfx_recv & achievement_recv
Event readers should be ordered after their respective writers within the frame. Undefined ordering between writers and readers can lead to subtle out of order bugs. Delaying communication across frames is often not intentionally desired. If it is something you want it should be made explicit.
There are exceptions for systems like achievements or analytics, but I'd only recommend excluding them from ordering if you have a good reason. Often they will not be computationally intense, so having them all run at the end of frame is fine.
You can achieve this by using event_producer.before(event_consumer) or (event_producer, event_consumer).chain() when adding systems for systems within the same SystemSet. For events that cross a SystemSet boundary this should be taken care of by the ordering of the SystemSets in your app.configure_sets() call.
Systems that only do work based on an event should have that as part of their run condition.
fn handle_player_level_up_event(mut events: EventReader<PlayerLevelUpEvent>) {
events.iter().for_each(|e| {
// ...
});
}
handle_player_level_up_event.run_if(on_event::<PlayerLevelUpEvent>())Write helper utilities for common operations
Tag entities with a cleanup Zero Sized Type (ZST) component. We can then add our cleanup utility system with our new cleanup component as the type. This creates a simple and consistent way to remove all entities marked with the component when transitioning or exiting certain states.
#[derive(Component)]
struct CleanupInGamePlayingExit;
fn cleanup_system<T: Component>(mut commands: Commands, q: Query<Entity, With<T>>) {
q.for_each(|e| {
commands.entity(e).despawn_recursive();
});
}
// When spawning entities
commands.spawn((
Name::new("projectile"),
CleanupInGamePlayingExit,
...
));
// Add to state transition
.add_systems(
OnExit(GameState::InGame),
cleanup_system::<CleanupInGamePlayingExit>,
)Credit to bevy cheatbook.
When working with queries and the Entity type, often you'll be matching on the outcome to exit early if the entity was not found.
The tedium of writing match expressions all over the place to return early can be avoided through a few simple macros. I've provided a one but you can imagine more variations based on the methods on Query.
Do be careful when using these, as opposed to the panicing methods like query.single(), these will silently return. This may be appropriate for your game, however it could also lead to bugs and unusual behaviour if they were supposed to succeed.
You could even make variations of these that return in release but panic in debug if that fits your use case.
fn print_window_size(windows: Query<&Window>) {
let window = get_single!(windows);
println!("Window Size: {}", window.resolution);
}
#[macro_export]
macro_rules! get_single {
($q:expr) => {
match $q.get_single() {
Ok(m) => m,
_ => return,
}
};
}Bevy utlises a prelude module to great effect for easy access to common imports. We can do the same!
By creating a prelude module in our project and exporting the various types that are commonly used within our game we can greatly cut down on the number of imports we need to maintain. You can also bring in the preludes from commonly used dependencies if you like. I have done that here with bevy and rand.
A nice side effect of this pattern is moving around code or refactoring doesn't require changes in as many places. If you restructure your audio code, you only need to update how it's presented in the prelude, assuming the rest of your project utilises the prelude.
src/audio.rs
pub(crate) mod prelude {
pub(crate) use super::{EventPlaySFX, SFXKind};
}
#[derive(Event)]
pub(crate) struct EventPlaySFX { /* ... */ }
pub(crate) enum SFXKind { /* ... */ }src/prelude.rs
pub(crate) use bevy::prelude::*;
pub(crate) use rand::prelude::*;
// Common items available at the root of the prelude
pub(crate) use crate::{Enemy, Health};
// Specific areas nested within their own module for self documenting use
pub(crate) mod audio {
pub(crate) use crate::audio::prelude::*;
}
pub(crate) mod physics { /* ... */ }src/enemy.rs
use crate::prelude::*;
fn handle_enemy_health_changed(
mut commands: Commands,
enemies: Query<(&Health, Entity), (With<Enemy>, Changed<Health>)>,
mut play_sfx: EventWriter<audio::EventPlaySFX>,
) {
enemies.for_each(|(health, id)| {
if health.current <= 0. {
commands.entity(id).despawn_recursive();
play_sfx.send(audio::EventPlaySFX::new(audio::SFXKind::EnemyDeath));
}
});
}Bevy Plugins enable grouping systems, components, and resources into logical units. They're used heavily in Bevy itself and are what powers the ability to turn on/off parts of the engine.
By constructing your game out of plugins you make it easier to find, work with, and debug subsystems. It also contextualises the setup and configuration of 3rd party crates to where they belong. For example, setting up the resources, plugins, and systems to utilise a 3rd party terrain library would go in your TerrainPlugin. That way, disabling your own terrain plugin will also disable the library you've imported, and any other resources that only it needed.
Note
Your mileage may vary with "enabling"/"disabling" plugins in your game. Bevy implements it in engine because it's valuable to disable chunks of the engine. However to achieve this in the game itself is not only more difficult, but the payoff is lower. How often will you realistically want to remove physics or audio from your game?
src/audio.rs
pub(crate) struct AudioPlugin;
impl Plugin for AudioPlugin {
fn build(&self, app: &mut App) {
app.add_plugins(some_audio_library::AudioFXPlugin)
.init_resource::<MyAudioSettings>()
.add_systems(...);
}
}src/physics.rs
pub(crate) struct PhysicsPlugin;
impl Plugin for PhysicsPlugin {
fn build(&self, app: &mut App) {
app.add_plugins(some_physics_library::BouncyPhysicsPlugin)
.init_resource::<MyPhysicsSettings>()
.add_systems(...);
}
}src/game.rs
pub(crate) struct GamePlugin;
impl Plugin for GamePlugin {
fn build(&self, app: &mut App) {
app.add_plugins((
DefaultPlugins,
crate::AudioPlugin,
crate::PhysicsPlugin,
));
}
}src/main.rs
fn main() {
bevy::prelude::App::new()
.add_plugins(crate::game::GamePlugin)
.run();
}Bevy's dependencies do a lot of trace logging that is not relevant for an end user.
To improve your runtime performance, you can add the following to the [dependencies] section of your Cargo.toml.
It will disable high log levels on compile time so that they do not need to be filtered out while your app is running.
log = { version = "0.4", features = ["max_level_debug", "release_max_level_warn"] }Bevy does not yet provide it's own build system, so we get the default Rust profiles of dev and release. While these profiles are a decent starting point, they're conservative and need to cater to a wide set of circumstances.
If you're following the Bevy getting started guide you'll already have encountered some of what we'll be doing.
Some of the build commands will require extra flags which can be annoying to type and easy to forget. I recommend some small layer infront of them. Anything would work here including batch files, shell scripts, makefiles, or something like just which is my preference.
The default Rust linker is often a significant portion of compile times. If you're using Bevys dynamic_linking described in the Development section this shouldn't be a big deal, however much faster linkers exist that you can experiment with such as LLD or mold. They are mentioned in the Bevy getting started guide.
TODO: Look into CPU features cross referenced with steam cpu feature support. See "Other Settings" at the bottom of Steam Hardware Survey.
Here are the settings I use for the fastest iteration build times. If you require more performance in debug mode you can try the suggestions in the code snippet. I've generally found going above opt-level = 2 doesn't give a meaningful increase for the compile time cost.
[profile.dev]
debug = 0
strip = "debuginfo"
opt-level = 0 # Switch to 1 for more runtime performance
# overflow-checks = false # Uncomment for better math performance
[profile.dev.package."*"]
opt-level = 2We disable certain debug information as it increases compile & link times. We are still able to get stack traces even with them disabled. You can always re-enable them if you have a specific need.
You'll also want to be using Bevy's dynamic_linking feature eg cargo run -F bevy/dynamic_linking. Dynamic linking allows us to avoid paying the cost of linking all of Bevy and it's dependencies every time we compile.
Our release profile doesn't look much different. release comes with opt-level = 3 by default but we'll specify it for clarity. We'll continue with removing debug info as we did in dev.
[profile.release]
opt-level = 3
panic = 'abort'
debug = 0
strip = "debuginfo"
# lto = "thin" # Enable for more inlining with a bigger tradeoff in compile timesThe main changes from dev are the optimisation level, aborting on panics, and no longer using the Bevy dynamic_linking feature. We don't need unwinding on panics as Rust favours tools like Result and being able to match and handle errors gracefully. Disabling unwinding reduces how much code is generated leading to better performance, more inlining opportunities, and smaller builds.
For a little more potential speed you can try "thin" link time optimisation, though I've found the increase in interative compile times not worth it.
Our distribution profile will be what we ship to players. We create it by specifying that it inherits from our release profile, then we'll tune a few more options.
[profile.distribution]
inherits = "release"
strip = true
lto = "thin"
codegen-units = 1strip = true will remove even more debug information, shrinking the size of the binary while making stack traces much less useful.
lto = "thin is a given here, allowing for inlining opportunities across your games code and all your dependencies. You absolutely do not want Bevy dynamic_linking enabled for distribution builds as it prevents propper LTO. There is also another lto option called "fat" however I would discourage it's use. It is constrained to a single core and will dramatically increase compile times while often having no benefit over thin.
By default rustc will split crates into a number of "code generation units" to allow for parallel compilation within each crate. codegen_units = 1 disables this feature, leading to generally faster code with longer compile times per crate.
Lastly, we will disable the logging capabilities of our build. End users will never see or need this information. Because it increases the size of the build, slows compile times, and hurts performance it's a good option unless you have very specific needs.
Unfortunately it's not as easy as any of the previous options. You will need to add the tracing and log to your Cargo project and ensure they're the same versions being used by Bevy. Make sure there isn't multiple entries in your Cargo.lock file. Then when building the game you'll need to specify features within both of them.
Your build command will look something like this.
cargo build --profile distribution -F tracing/release_max_level_off -F log/release_max_level_offGenerally the cost of each level of logging increases as you go from error to trace, so you might opt for keeping error logs as they're usually quite infrequent and often not in the "hot path". In that case you can switch from release_max_level_off to release_max_level_error.
Except where noted, all code in this repository is dual-licensed under either:
- MIT License (LICENSE-MIT or http://opensource.org/licenses/MIT)
- Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
at your option. This means you can select the license you prefer! This dual-licensing approach is the de-facto standard in the Rust ecosystem and there are very good reasons to include both.
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.